I recently noticed that Blue Cat’s Plug’n Script makes a couple of appearances in Jim Lill’s latest video exploring microphone preamps and in particular Neve consoles’ preamps. For those who …

Stop thinking of delay as an echo. Start thinking of it as a creative system for building entirely new sounds and getting inspired instead. Most delay plug-ins are built to …

WWDC answered whether your assistant is private. It never answered whether it’s telling the truth — and Apple just gave it hands.

The smartest thing I’ve read about Apple’s WWDC didn’t come from Apple. It came from an analyst named Nate B. Jones, who watched the same keynote everyone else did and noticed that the real story wasn’t whether Siri had finally gotten smart. The real story, he argued, is a land grab over what he calls the trusted action surface — the place where AI actually meets your work, touches your apps, and is handed permission to do something. There are two great bottlenecks in AI, he points out: raw compute, which is Jensen Huang’s kingdom, and the trusted surface where intelligence becomes useful, which is the one Apple just went to war for. Whoever owns that surface owns the meter when intelligence becomes unavoidable. It’s a sharp frame, and he’s right.

He’s right, but he left out the scariest part.

Here is what Apple actually did. It tore Siri down to the studs and rebuilt it on Google’s Gemini — reportedly a custom, 1.2-trillion-parameter model that Apple pays Google something on the order of a billion dollars a year to use. It gave the new assistant eyes: real-time awareness of what’s on your screen. And — this is the part that matters — it gave it hands. The new Siri doesn’t just answer. It manages your browser tabs, rewrites your weak passwords, reaches across your apps, pulls context out of Mail and Messages in the middle of a phone call, and acts inside the software where your life actually happens. Craig Federighi wrapped the whole thing in a promise about privacy and took a quiet shot at the rest of the industry for chasing AI for its own sake while losing sight of the people it’s supposed to serve.

It was a good keynote. And it answered exactly one of the two questions that matter.

Apple answered is it private? — will the assistant know your life without strip-mining it and selling the tailings? That’s a real question, and Apple has a real, earned answer. The other question, the one nobody on that stage went near, is is it true? And the instant you give an assistant hands, that second question stops being academic.

Think about what changes when an assistant goes from talking to doing. When a chatbot makes something up, it costs you a minute and a raised eyebrow. When an agent makes something up and then acts on it, it sends the email to the wrong person, moves money to the wrong account, deletes the file it meant to keep, books the flight for the wrong Thursday. A hallucination in a chat window is an annoyance. A hallucination with hands is an incident. Agency multiplies the cost of being wrong by exactly the thing that makes agency valuable.

And what is driving those hands? A large language model — a very capable one, but the same kind of machine that, like every model of its line, fabricates with total confidence when it doesn’t know. Gemini is excellent and Gemini hallucinates; both are true, the way both are true of all of them. Apple took the most capable probabilistic guess-engine it could license, gave it the keys to your apps and permission to act, and then reassured you about privacy. The locks on the doors are magnificent. Nobody mentioned whether the butler tells the truth.

This is the gap I’ve spent this whole series circling, and I’ll disclose again that I co-founded a company built on closing it, so weigh that however you like. But the principle stands on its own, and it is bigger than any one company: a trusted action surface is only as trustworthy as the facts the agent acts on. You can own the device, the operating system, the permission prompt, the whole beautiful surface — and if the thing deciding what to do is guessing, you have built a faster, smoother way to be confidently wrong about someone’s money. The surface needs a substrate. The hands need a conscience. Trust has two axes — privacy and veracity — and at WWDC Apple shipped one of them and didn’t mention the other was missing.

Which loops back to Jensen, and to the argument I’ve been making from the other direction. Nate’s case is that value migrates off the model and onto the surface, and that this is NVIDIA’s problem. Mine has been that value migrates off the GPU and onto a humbler kind of silicon — the CPU —, because most of what we ask AI to do is look something up, not dream. Two different roads, one destination: the belief that NVIDIA’s position is a law of nature is a story, not a fact. But notice that the surface only wins if people trust it — and the agentic surface raises the trust bar at the very instant it raises the stakes. Owning the meter is worthless if the meter lies.

So watch the surfaces, as Nate says. But watch what they’re built on. Apple just gave a billion people’s computers hands, eyes, and access, and wrapped it in the best privacy story in the industry. It gave Siri hands before it gave Siri honesty. Until the conscience ships, the hands are the part I would keep an eye on.

Robert X. Cringely is a co-founder of 2Brains, Inc., in Charlottesville, Virginia. He has written this column since 1987.

The post Apple Gave Siri Hands first appeared on I, Cringely.






Digital Branding
Web Design Marketing

Why the machines hallucinate, why they have no worldview, and why truth has to come from somewhere else.

I’m going to say something that sounds like an insult and is meant as a description: large language models (all of them) hav never known a true thing. Not once. It doesn’t know things at all. It is extraordinarily good at sounding like it does, which is a different skill, and most of our present confusion comes from mistaking the second for the first.

Here is what a language model actually does. It has read an enormous amount of text, and from that text it has learned, with real brilliance, what tends to come next. Give it some words and it predicts the words likely to follow. That is the whole trick. It is a magnificent trick — it gives us machines that write fluent prose in any voice on any subject — but look at what it optimizes for. It optimizes for plausible. It was never, at any point, optimizing for true. Truth was not in the objective. Plausibility was. And plausibility and truth often travel together, which is precisely why we confuse them — but they are not the same thing, and the gap between them is the whole story.

This is why these systems “hallucinate,” a word I dislike because it implies a malfunction. There is no malfunction. A model that invents a court case that never happened — complete with a docket number, plausible parties, and a tidy holding — is not broken. It is doing exactly what it was built to do: produce the most plausible continuation. A fake citation is plausible. It looks like the thousands of real ones the model has read. The machine has no way to prefer the real one, because it has no idea that “real” is a category. It isn’t lying, either. Lying requires knowing the truth and choosing against it, and the machine has never once been in a position to know.

Now the deeper point, the one that took me a long time to learn to say cleanly. Truth is not a property of language. You cannot find it inside a sentence by examining the sentence harder. Truth is a property of the relationship between a sentence and the world — between the words “it is raining” and the actual sky. A statement is true when it corresponds to how things are. And the model has only ever seen the words. It has read every description of rain ever written and stood out in none of it. It holds the map — all of the maps, every map anyone has ever drawn — and it has never once been to the territory. That is why it can be eloquent and wrong in the same breath and feel no friction between the two. The friction lives in a place the model has never visited.

There’s a corollary that unsettles people, and it shouldn’t. A machine like this has no worldview. None. It will argue any side of anything with equal grace, defend a position and then dismantle it in the next window, because it isn’t holding a position — it’s rendering one. It is a mirror with a vocabulary. We keep waiting for it to reveal what it really believes, and it doesn’t believe anything, and that is not a flaw to be trained out of it. It is the honest fact of the thing. The language is separate from any view of the world. That was the original insight some of us started from years ago, before any of the building began: language is machinery, and machinery has no creed.

It is a mirror with vocabulary

The trouble is that we keep dressing the machinery in the costume of a knower. We put it behind a chat window that answers in the first person, warm and certain, and every instinct we have says this thing believes what it is telling me. It does not. It cannot. And the distance between how it sounds and what it is happens to be the most dangerous real estate in the whole technology, because that is exactly where a fluent falsehood gets received as a considered judgment — in a clinic, in a courtroom, in a loan decision, in a room where someone is deciding whether to act.

So what do you do with a machine that can say anything and stand behind nothing? You stop asking it to be the thing it cannot be. If truth lives in the relationship between a claim and the world, then truth has to come from the world — from some grounded, checkable account that sits outside the language model and stays outside it. You don’t teach the renderer to be honest. You keep the saying and the knowing in separate rooms, and you let the language render only what the knowing will vouch for. Language on one side, a verifiable account of the world on the other, and a wall between them you can actually inspect.

That sounds tidy until you try to build it, and then you hit the part nobody puts on a slide. Before you can check a claim against the world, you have to know what the claim is — and pulling discrete, checkable claims out of fluent prose is genuinely hard. The machine doesn’t speak in clean facts. It speaks in paragraphs, where an assertion hides inside a subordinate clause, where a hedge can pass for a claim and a claim can pass for a hedge, and where — my favorite trap — every individual sentence is true and the paragraph they assemble into is a lie. The honest sentence, marshaled into a dishonest whole. Working out what is actually being asserted, before you have checked whether any of it is so, turns out to be most of the labor. It is unglamorous, and it is the ballgame.

I don’t think the future of this technology is a more fluent machine. We already have fluency. Fluent is solved. The future is a more honest architecture — one that knows the difference between what it can say and what it can stand behind, and that keeps the truth somewhere you can point to and check. A machine with no worldview is not the problem. Pretending it has one is. The repair was never going to be giving the machine a conscience. It is to stop asking the part that talks to also be the part that knows.

Full disclosure: I’m a co-founder of 2Brains, a company built on exactly this conviction, so I am not a neutral party here, which we have solved and have patent pending. But the conviction came first. The company exists because of it, not the other way around.

 

The post GenAI is Fluent in Everything, but Faithful in Nothing first appeared on I, Cringely.






Digital Branding
Web Design Marketing

Every few weeks, someone announces a tool that detects AI hallucinations. A startup, a research lab, a hyperscaler bolting a “trust layer” onto its chatbot. The release uses the word “guardrails.” Everyone nods. Another brick in the road to safe, reliable AI.

I want to argue that we are cheering for the wrong thing — that hallucination detection, however clever, cannot be the strategy. It can be a backstop. It can be a monitor. It cannot be the plan. And the reason is older than computing.

Start with the trap at the center of the whole idea.

To catch a hallucination, your detector has to know the right answer. Sit with what that means. The original model produced a confident falsehood because it did not have the grounded knowledge to do otherwise. Now you propose a second system to sit behind it and flag the lies. But to flag a lie, that second system has to know the truth — and if it knew the truth, you would not have needed the first model to guess in the first place. You would just serve the truth and skip the theater.

A detector good enough to reliably catch fabrication would have to possess exactly the capability whose absence caused the fabrication. Detection doesn’t solve the problem. It assumes the problem is already solved. That is the whole argument in a paragraph; everything else is just watching it play out.

So watch it play out. The first thing you notice is that a hallucination has no tell. When one of these models invents a court case, a citation, a drug dosage, a quarterly number, the sentence it produces is grammatically perfect, tonally identical to a true one, and delivered with precisely the same confidence. The model is not more hesitant when it lies. It does not sweat. There is no flicker. That is the entire reason this is hard: the false output and the true output are indistinguishable on their face. A detector staring at the text has nothing to grab onto, because there is nothing in the text to grab.

So the detector-builders do the sensible thing and go probabilistic. They get good — let’s be generous and say 95% good. And 95% sounds like an A. But invert it. In a hospital, a courtroom, a bank, a grid control room, 95% means one in twenty confident falsehoods walks right past the guard. And here is the cruel part: the ones that get through are not random. They are the most plausible fabrications in the batch — the ones convincing enough to fool the detector, which makes them precisely the ones most likely to fool you. A safety system that is only probabilistic is not a safety system. It is a liability with a press release.

It is also a treadmill. Every new model, every new domain, every fresh way of being wrong demands that the detector be retrained and re-tuned. It is antivirus software for an attacker that rewrites itself weekly — perpetual catch-up, by design. And you pay for it twice: once to generate the answer, again to check it, and you still don’t get certainty for the money.

But the deepest mistake here is a category error, and to name it I have to wade back into a fight I picked a quarter century ago.

Everyone reaches for W. Edwards Deming when they talk about quality — the American sage the Japanese supposedly heeded when Detroit wouldn’t. I once spent 4,400 words arguing the standard story gets the hero wrong. The man who actually carried disciplined quality into occupied Japan was a 29-year-old radio engineer named Homer Sarasohn, sent by MacArthur in 1946 to rebuild a flattened electronics industry. He and his colleague Charles Protzman, a Western Electric production man, spent four years teaching Japanese executives how to run a company and build things that worked — they literally wrote the handbook for it, a course book still in print in Japan today — and when they went home, Sarasohn handed the baton to Deming, who had a gift for self-promotion and ended up with his name on the prize and the legend. (Sarasohn was no footnote; he went on to a long career at IBM. History simply looked past him.) A remarkable number of readers wrote in to tell me I had it backwards. I didn’t, and I still don’t.

When that column ran, the Deming faithful came for me. The real transformation, they insisted, came from a handful of lectures Deming gave Japanese executives in the summer of 1950 — as if quality had arrived by seminar. Nonsense. If a few brilliant talks were all it took, answer me this: why did it take the better part of thirty years for Japan to turn quality into a weapon? The tools had been on the shelf since 1950 — Sarasohn’s manual, Protzman’s production discipline, Deming’s statistics, all of it.

What finally lit the fire was the memory chip. When Hitachi and the other Japanese makers went after the DRAM business Intel had invented, they slammed into the cruelest arithmetic in manufacturing: in a commodity chip, yield is the entire margin — and theirs was too low to make a dime. The answer had been sitting in Sarasohn’s handbook for three decades: build quality into the process instead of inspecting the failures out at the end. This time they used it. Japanese yields climbed past the Americans’ — seventy and eighty percent against Intel’s fifty or sixty — and by the mid-1980s the company that invented the DRAM had been driven out of it. The instruction was never the bottleneck. Necessity was.

We just prefer the story where one clever intervention saves the day — which is exactly the story being sold to us again: that a hallucination detector will do for AI what we like to pretend a seminar did for Japan.

But here is what matters for our purposes, and it is bigger than who gets the statue. Whether you credit Sarasohn, Deming, or the Japanese engineers who did the actual work, they all arrived at the same unglamorous law: you cannot inspect quality into a product. Sarasohn found factories where “quality” meant building a pile of vacuum tubes and throwing ninety percent of them away — where no one saw the problem with assembling precision electronics in a shack with a dirt floor. You do not fix that by hiring more inspectors to stand at the end of the line catching the bad ones. Inspection is expensive, it is late, and it never catches everything. The only thing that works is to build quality in — to design the process so the defect never happens. The industry that learned this went on to bury the one that had won the war. We are still driving the proof.

Hallucination detection is the man with the clipboard at the end of the line. It is quality by inspection, in a field that should have learned the lesson from manufacturing forty years ago.

And here is the part the clipboard can never fix: hallucination is not a malfunction. The model isn’t breaking when it makes things up. It is doing exactly what it was built to do — predict the most plausible next word, with no native notion of whether that word is true. Fabrication isn’t a bug in the architecture. It is the architecture, working as designed. You cannot detect your way out of a feature.

Which points at the only strategy that survives contact with the problem. Stop trying to catch the lie after the fact, and build a system that knows the boundary of what it actually knows — one that can tell the difference between answering from grounded, verified knowledge and reaching past the edge into invention, and that says so when it gets there. Not a smarter smoke detector. A machine that doesn’t set the fire.

That is harder. It is architectural, not bolted on, and it does not make for a tidy press release about a new trust layer. But it is the only version of this that works in a courtroom, where “our filter catches 95%” is not a sentence you want to say to a judge.

Detection is not a strategy. Design is. Sarasohn knew it in 1948. It is past time we learned it about machines that talk.

(Disclosure: I co-founded 2Brains, which is built around designing it in rather than inspecting it out, so I come to this with a horse in the race. I’d make the argument anyway — I was making versions of it about Japanese factory floors a quarter century ago.)

The post Detection Is Not a Strategy first appeared on I, Cringely.






Digital Branding
Web Design Marketing

404 Media remembers how a Florida police office looked up his ex-girlfriend's license plate in the Flock automated license plate reader system at least 69 times in 2024 — even searching for her mom's license plate at least 24 times. The police office was charged with stalking and hacking-related offenses, serving one day in prison with five years of probation — but his case "was not a one-off." [Alternate link via Bruce Schneier] Local news reports from around the country repeatedly detail police abusing the Flock surveillance system in order to stalk their partners or ex-partners. The contours of each story are much the same, with the police officer in question using their access to the system to repeatedly track a specific person over the course of weeks or months. The cases highlight the fact that Flock can be used to track the whereabouts of individual people, that police do not get a warrant in order to use the system, and that, if they have access to the system, they have the technical ability to look up any license plate they want for any reason they want. An April study by the civil rights group Institute for Justice found that at least 18 police officers have been caught around the country using Flock to stalk a romantic interest in the last few years; another database, called the ALPR Abuse Library, has documented 20 specific cases of "stalking/targeting" around the country. The known cases of police stalking are almost certainly a vast underreporting of the overall abuse, because they largely include only cases in which the behavior was so egregious that it led to police officers being fired, arrested, or both. Flock told 404 Media that it is "aware of 15 incidents of abuse, each surfaced because of the transparency and accountability features deliberately built into our platform.... There are also 140,000 monthly active users of Flock, so the relatively rare instances of abuse, while obviously wrong and awful, are exactly that — rare," a Flock spokesperson told 404 Media. [One in 10,000.] "Humans are fallible; unlike most tools society provide law enforcement, Flock ensures that in the instances when our technology is misused, the evidence used to hold responsible parties accountable, is right there in our system. We also encourage all our customers to have a usage policy, regular training, and to implement our Audit Assistance tool, which proactively flags unintended use...." But it is also the case that Flock has strenuously fought against lawsuits and potential regulations that are seeking to require police to get a warrant to use the system. And many cases of abuse have not been detected by police departments themselves but by those private citizens, journalists, and stalking victims who have found patterns of abuse in public records files they have obtained from their local police departments. In most cases of Flock-related stalking reviewed by 404 Media, the abuse occurred over the course of months or years, and the victims were subjected to dozens or hundreds of lookups. Other abuse cases have been discovered using the website HaveIBeenFlocked.com, a website that compiles Flock searches released via public records requests and turns them into a searchable database. Flock has repeatedly tried to get that website taken down, as we have previously reported.

Read more of this story at Slashdot.